Three-Dimensional Point Cloud Model Reconstruction Method, Computer Readable Storage Medium and Device

1. A three-dimensional point cloud model reconstruction method, comprising: 1) sampling and WLOP-consolidating an input point set to generate an initial surface point set, copying the initial surface point set as an initial position of an interior skeleton point set, to establish a correspondence relation between surface points and skeleton points; 2) moving points in the interior skeleton point set inwards along a direction opposite to a normal vector thereof, to generate interior points; 3) using a self-adaptive anisotropic neighborhood as a regularization term to perform an optimization of the interior points, and generating skeleton points; 4) performing a consolidation and completion of the initial surface point set using the skeleton points, to generate consolidated surface points; 5) reconstructing a three-dimensional point cloud model according to the skeleton points, the surface points and the correspondence relation between the surface points and the skeleton points; wherein establishing a correspondence relation between surface points and skeleton points in step 1) comprises: constructing a deep point set <P,Q>={<p i ,q i >} i∈I ⊂R 6 according to the surface point set and the skeleton point set, a deep point in the deep point set being composed of a point pair <p i ,q i >, wherein p i s a point in the surface point set P={p i } i∈I ⊂R 3 , q i is a point in the skeleton point set Q={q i } i∈I ⊂R 3 , and I is a sampled point set; a direction of a deep point pair is m i =(p i −q i )/∥p i −q i ∥ and consistent with a normal vector of the surface point; between step 1) and step 2), further comprising: determining a size of a neighborhood of each point in the surface point set and the interior skeleton point set, wherein, the neighborhood of the surface point is P ~ = { p i ′ |  p i ′ - p i  < σ p ⁢ r } , wherein a default value of σ p is 5; the neighborhood of the interior skeleton point is Q ~ i = { q i ′ |  q i ′ - q i  < σ q ⁢ r } , wherein a default value of σ q is 2; r is an average distance r = 1  P  ⁢ ∑ i ∈ I ⁢ min i ∈ I ⁢ \ ⁢ { i } ⁢  p i - p i ′  between each sample point and an adjacent point; |P| is the number of the surface points; p i is a point in the surface point set, and p i′ is a surface point adjacent to p i .

2. The three-dimensional point cloud model reconstruction method according to claim 1 , wherein moving points in the interior skeleton point set inwards along a direction opposite to a normal vector thereof, to generate interior points in step 2) comprises: determining a condition for stopping the inward movement of each sample point in the interior skeleton point set, wherein the condition for stopping the inward movement of the sample point is that in a neighborhood {tilde over (Q)} of the sample point, the maximum angle between the normal vector of sample point and those of neighboring points is smaller than a preset threshold ω , i . e . , max i ′ ∈ I i Q ⁢ n i ′ · n i ≤ cos ⁡ ( ω ) wherein a default value of ω is 45°; moving each sample point along a direction opposite to a normal vector thereof according to the determined condition for stopping the inward movement of the sample point; after each step of movement of each point, a bilateral smoothing is performed for the current point to determine that q i = ∑ i ′ ∈ I i p ⁢ θ ⁡ ( p i , p i ′ ) ⁢ ϕ ⁡ ( n i , n i ′ ) ⁢ q i ′ ∑ i ′ ∈ I i p ⁢ θ ⁡ ( p i , p i ′ ) ⁢ ϕ ⁡ ( n i , n i ′ ) , wherein θ ⁡ ( p i , p i ′ ) = e - (  p i - p i ′  i ′ ) 2 , ϕ ⁡ ( n i , n i ′ ) = e - ( 1 - n i T ⁢ n i ′ 1 - c ⁢ ⁢ o ⁢ ⁢ s ⁡ ( ω ) ) 2 , I i p is a surface point set, n i is a normal vector of the sample point, and n i′ is a normal vector of an adjacent point.

3. The three-dimensional point cloud model reconstruction method according to claim 2 , wherein moving each sample point along a direction opposite to a normal vector thereof further comprises: an initial moving step length t of each sample point is r/2, wherein r is an average distance between each sample point and a closest adjacent point; a subsequent moving step length of each sample point is an average value of the moving step lengths of its adjacent point in a previous iteration, and a dynamic neighborhood Q i = { q i ′ |  q i ′ - q i  < σ q ⁢ r } ⁢ ⁢ is ⁢ ⁢ used .

4. The three-dimensional point cloud model reconstruction method according to claim 3 , wherein using a self-adaptive anisotropic neighborhood as a regularization term to perform an optimization of the interior points, and generating skeleton points in step 3) comprises: making a principal component analysis (PCA) for the interior point set to determine three principal axes and lengths thereof; determining the regularization term according to the determined principal axes and lengths thereof, wherein, R ⁡ ( Q ) = ∑ i ∈ I ⁢ λ i ⁢ ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ ϑ ⁡ ( q i , q i ′ )  d e ⁡ ( q i , q i ′ )  3 d e ⁡ ( q i , h k ) =  Å i T ⁡ ( q i - h k )  , ⁢ Å i = [ v i 1 / l i 1 ; v i 2 / l i 2 ; v i 3 / l i 3 ] , v i 1 , v i 2 , v i 3 are three determined principal axes, and l i 1 , l i 2 , l i 3 are the lengths of the three principal axes; analyzing an optimization arg ⁢ ⁢ min Q ⁢ ∑ i ∈ I ⁢ ∑ k ∈ I ⁢ ϑ ⁡ ( q i , h k ) ⁢  q i - h k  + R ⁡ ( Q ) according to the determined regularization term to generate the skeleton points; wherein, ϑ ⁡ ( q i , h k ) = e - d e 2 ⁡ ( q i , h k ) / r 2 is the self-adaptive anisotropic neighborhood, h k is a stationary point in the interior point set H={h i } i∈I , Q is a skeleton point set generated after the consolidation, q i is a point in the skeleton point set generated after the consolidation, and R(Q) is the regularization term.

5. The three-dimensional point cloud model reconstruction method according to claim 4 , wherein an analytic expression of the skeleton point in the skeleton point set is: q i = ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ α ii ′ ⁢ q i ′ ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ α ii ′ + μ ⁢  l i  2 ⁢ ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ β ii ′ ⁢ Å i T ⁢ Å i ⁡ ( q i - q i ′ ) ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ β ii ′ wherein α ii ′ = ϑ ⁡ ( q i , q i ′ )  q i - q i ′  , ⁢ β ii ′ = ϑ ⁡ ( q i , q i ′ )  d e ⁡ ( q i , q i ′ )  5 , ⁢ μ ⁢  l i  2 = λ i ⁢ ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ β ii ′ / ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ α ii ′ , d e ⁡ ( q i , q i ′ ) =  A ∘ i T ⁡ ( q i - q i ′ )  , l i = [ l i 1 ; l i 2 ; l i 3 ] , μ = 0.4 , and ⁢ ⁢ q i ′ is a point closest to the skeleton point q i .

6. The three-dimensional point cloud model reconstruction method according to claim 5 , wherein performing a consolidation and completion of the initial surface point set using the skeleton points, to generate consolidated surface points comprises: solving optimization arg ⁢ ⁢ min P ⁢ ∑ i ∈ I ⁢ η ⁡ ( p i ) ⁢ ∑ c j ⁢ ∈ i ⁢ θ ⁡ ( p i , c j ) ⁢  ( p i - c j ) ⁢ n i T  + R ^ ⁡ ( P ) + G ⁡ ( P ) to perform the consolidation and completion of the initial surface point set; wherein, η(p i ) is a density function, η ⁡ ( p i ) = 1 + ∑ c j ∈ ⁢ i ⁢ θ ⁡ ( c j , p i ) , θ ⁢ ( c j , p i ) = e - (  c j - p i  r ) 2 , and ⁢ ⁢ r = 1  P  ⁢ ∑ i ∈ I ⁢ min i ∈ I ⁢ \ ⁢ { i } ⁢  p i - p i ′  ; {circumflex over (R)}(P) is a regularization term, R ^ ⁡ ( P ) = ∑ i ∈ I ⁢ ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ θ ⁡ ( p i , p i ′ )  B i T ⁡ ( p i - p i ′ )  3 , θ ⁡ ( p i , p i ′ ) = e - (  p i - p i ′  r ) 2 , and ⁢ ⁢ B i = [ u i 1 ; u i 2 ] , which are any two orthogonal vectors on a tangent plane perpendicular to the normal vector; η i is a normal vector of the surface point, G(P) is a volume preservation term determined according to the skeleton point, q i is a point in the determined skeleton point set, and p i is a point in the initial surface point set.

7. The three-dimensional point cloud model reconstruction method according to claim 6 , wherein the volume preservation term is G ⁡ ( P ) = 1 2 ⁢ ∑ i ∈ I ⁢ γ ⁡ ( p i ) ⁢ (  p i - q i  - L ⁡ ( p i ) ) 2 , wherein , γ(p i ) is a weighting function, γ ⁡ ( p i ) = ( 1 + var ⁡ ( {  p i ′ - q i ′  } i ′ ∈ I i P ) ) - 1 , L ⁡ ( p i ) = ∑ i ′ ∈ I i P ⁢ θ ⁡ ( p i , p i ′ ) ⁢  p i ′ - q i ′  ∑ i ′ ∈ I i P ⁢ θ ⁡ ( p i , p i ′ ) , L ⁡ ( p i ) is an evaluated volume thickness of the current point determined according to the skeleton point, q i is a point in the determined skeleton point set, and p i is a point in the initial surface point set.

8. The three-dimensional point cloud model reconstruction method according to claim 7 , wherein the consolidated surface point is p i = η ⁡ ( p i ) ⁢ ∑ c j ⁢ ∈ i ⁢ α ij ⁡ ( c j - p i ) ⁢ n i T γ ⁡ ( p i ) + η ⁡ ( p i ) ⁢ ∑ c j ⁢ ∈ i ⁢ α ij + γ ⁡ ( p i ) ⁢ ( q i + L ⁡ ( p i ) ⁢ m i ) γ ⁡ ( p i ) + η ⁡ ( p i ) ⁢ ∑ c j ⁢ ∈ i ⁢ α ij + μ ^ ⁢ ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ β ^ ii ′ ⁢ A ∘ i T ⁢ A ∘ i ⁡ ( p i - p i ′ ) ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ β ^ ii ′ ⁢ wherein , α ij = θ ⁡ ( p i , c j )  p i - c j  , β ^ ii ′ = θ ⁡ ( p i , p i ′ )  B i T ⁡ ( p i - p i ′ )  5 , μ ^ = ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ β ^ ii ′ / ( γ ⁡ ( p i ) + η ⁡ ( p i ) ⁢ ∑ c j ⁢ ∈ i ⁢ α ij ) , B i = [ u i ∼ 1 ; u i ∼ 2 ] , which are any two orthogonal vectors on a tangent plane perpendicular to the normal vector; m i is a direction of a deep point pair; i′ is an adjacent point of the sample point.

9. The three-dimensional point cloud model reconstruction method according to claim 8 , further comprising, between step 4) and step 5): correcting the normal vector obtained through the PCA calculation using the orientation of the deep point, and a distribution condition of the input points around the deep point, i.e., n i = ( η ⁡ ( p i ) - 1 ) ⁢ n ∼ i + m i  ( η ⁡ ( p i ) - 1 ) ⁢ n ∼ i + m i  .

10. The three-dimensional point cloud model reconstruction method according to claim 6 , wherein the volume preservation term is G ⁡ ( P ) = 1 2 ⁢ ∑ i ∈ I ⁢ γ ⁡ ( p i ) ⁢ ( L ⁡ ( p i ) ) 2 , γ(p i ) is a weighting function, γ ⁡ ( p i ) = ( 1 + var ⁡ ( {  p i ′ - q i ′  } i ′ ∈ I i P ) ) - 1 , L ⁡ ( p i ) = ∑ i ′ ∈ I i P ⁢ θ ⁡ ( p i , p i ′ ) ⁢  p i ′ - q i ′  ∑ i ′ ∈ I i P ⁢ θ ⁡ ( p i , p i ′ ) , L ⁡ ( p i ) is an evaluated volume thickness of the current point determined according to the skeleton point, q i is a point in the determined skeleton point set, and p i is a point in the initial surface point set.

11. The three-dimensional point cloud model reconstruction method according to claim 10 , wherein the consolidated surface point is p i = η ⁡ ( p i ) ⁢ ∑ c j ⁢ ∈ i ⁢ α ij ⁡ ( c j - p i ) ⁢ n i T γ ⁡ ( p i ) + η ⁡ ( p i ) ⁢ ∑ c j ⁢ ∈ i ⁢ α ij + γ ⁡ ( p i ) ⁢ ( q i + L ⁡ ( p i ) ⁢ m i ) γ ⁡ ( p i ) + η ⁡ ( p i ) ⁢ ∑ c j ⁢ ∈ i ⁢ α ij + μ ^ ⁢ ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ β ^ ii ′ ⁢ A ∘ i T ⁢ A ∘ i ⁡ ( p i - p i ′ ) ∑ i ′ ∈ I ⁢ \ ⁢ { i } ⁢ β ^ ii ′ wherein , α ij = θ ⁡ ( p i , c j )  p i - c j  , β ^ ii ′ = θ ⁡ ( p i , p i ′ )  B i T ⁡ ( p i - p i ′ )  5 , μ ^ = ∑ i ′ ∈ I ⁢ { i } ⁢ β ^ ii ′ / ( γ ⁡ ( p i ) + η ⁡ ( p i ) ⁢ ∑ c j ⁢ ∈ i ⁢ α ij ) , B i = [ u i ∼ 1 ; u i ∼ 2 ] , which are any two orthogonal vectors on a tangent plane perpendicular to the normal vector; m i is a direction of a deep point pair; i′ is an adjacent point of the sample point.

12. A device, comprising: a scanning module for scanning a model to acquire an input point set; a memory for storing program instructions; and a processor connected to the scanning module and the memory, for executing the program instructions in the memory, and processing the input point set according to the steps in claim 1 .